Beyond Coding: The Native-Multimodal Debate Is Still Alive

Over the past year, Coding and Agent capabilities have rewritten the LLM leaderboards — they are also the fastest path to commercial value for AI. But beside that race, another, longer-running question is back on the table: should general-purpose LLMs be given eyes?

July 2026 saw Moonshot release Kimi K3 — a 2.8-trillion-parameter MoE with a 1-million-token context window, continuing the early-fusion recipe from K2.5. Visual and text tokens are mixed at a fixed ratio from the very start of pre-training, and the visual encoder — MoonViT-V2 (~400M parameters) — is trained from scratch rather than being initialized from SigLIP. On the Arena Frontend Code leaderboard, K3 hit 1679 points at one point. The browser-platform company Puter deliberately planted 5 visual discrepancies in a test page; K3 compared the target page with the rendered output and found all five, with zero false positives. Moonshot calls this loop "vision in the loop": the model writes code, inspects the result, and adjusts.

Around the same time, DeepSeek took a different route. The GA version of DeepSeek V4-Flash has 284B total parameters and activates only 13B per token — roughly one-tenth and one-eighth of K3's figures, respectively. No architecture change, no parameter inflation — just a sharp post-training push. The payoff shows up on leaderboards: V4-Flash (formal) substantially beat its own preview and even the V4-Pro preview across DeepSeek's reported Coding and Agent evaluations, and its Arena WebDev public score briefly climbed to 1577, putting it within reach of GLM-5.2.

The short version: on the same Coding table, K3 bets on giving the model real visual perception, while V4-Flash bets on grinding Coding scores to their absolute ceiling with much less compute.

Why Agents Must See

One multimodal researcher puts the argument simply: "On long-horizon tasks, errors compound when feedback only comes from the code path. Vision is a more accurate signal and a closer match for user intent."

Pure-text LLMs are still "blind." Real systems can wire in OCR, an independent VLM, or vision tools over MCP — converting images into text, tags, coordinates, or structured fields before handing them to the main model. That "add-on" path is perfectly workable in production, but the native-multimodal camp argues it has two ceilings: the image is compressed into text first, so the bandwidth narrows; and pulling visual feedback into a single training trajectory during post-training is much harder.

The same disagreement shows up at the field level. Yann LeCun's view: "the vast majority of human knowledge isn't expressed as text." Image, video, and real-world interaction remain the source from which any model that wants to understand objects, space, and action eventually has to learn. Ilya Sutskever's opposing take: "text is a projection of the world," and a model that keeps learning from text can still build world understanding. Both sides agree on the long-run value of multimodality. The disagreement is about when and at what cost.

The Cost of Multimodality: Not 1+1, but >2

Adding vision to an LLM looks like adding another input port. Underneath, it is a structural collision. Images and text have different data structures, information densities, and learning rates. Sharing one backbone forces them to fight for capacity against code, math, reasoning and each other, and training objectives interfere.

Kimi K2.5's technical report gives a concrete ablation: when visual data is added late in training, the model's text capability drops first and only slowly recovers. In other words, "1+1 > 2" is a real cost — the better vision-in-the-loop works, the more carefully text/Coding stability has to be rebalanced.

Two hidden costs compound that: compute and data. Apple's MM1 paper showed that visual encoder choice, image resolution, and the visual-token count all materially affect results; higher resolutions and more visual tokens mean higher training and inference cost. For tasks that only need to read a few fields, asking a general model to keep scanning the whole image is not obviously cheaper than running OCR or a specialist model.

China's Top Tier Has Already Split

By August 2026, the Chinese general-purpose foundation-model camp splits into two visible tracks:

  • Native-multimodal camp: Moonshot (Kimi K3, 2.8T / 104B activated), Alibaba (Qwen3.8-Max, 2.4T / 95B activated, combining native vision, Coding and Cowork), ByteDance (Doubao-Seed-2.1, with visual understanding and multimodal input as headline features) — all "bucket models" that natively fuse vision into very large parameter budgets.
  • Pure-text post-training camp: DeepSeek (V4-Flash at 284B / 13B), Zhipu (next-generation general base model expected to move further toward multimodality), Tencent HunYuan — text-in, text-out, with Coding/Agent scores optimized first.

DeepSeek founder Liang Wenfeng once said: "To train AI well, you don't need a world model — you don't even need multimodality." That line is often quoted as 'DeepSeek is bearish on multimodality.' It is less often quoted with the second half: "But multimodality is something we'll eventually have to do." Both halves are true. The real split is this: while Coding is still iterating fast, how much model capacity, data, and compute do you spend on training vision at the same time?

Two Clocks Are Racing

In the short run, Coding and Agent rankings are reshuffled every few weeks or months — that is one clock. In the long run, going from text-only LLMs to genuinely vision-grounded understanding is a multi-year marathon — that is another.

K3 is racing the first clock — using 2.8T parameters, native vision, and a 1M context to push Arena Frontend Code's ceiling higher. V4-Flash is also racing the first clock — using one-tenth the parameter count and post-training distillation to overtake a stack of previews on Arena WebDev. Both strategies work right now. But the divergence shapes how far each can eventually go:

  • When vision-in-the-loop becomes the default feedback channel for Coding Agents, the pure-text + post-training line will hit a "see / don't see" ceiling.
  • When the next multimodal benchmark redefines SOTA, the models that baked vision into the backbone earlier will start from a structurally stronger position.

So What

This is not a disagreement about the long-run AGI path for Chinese foundation models. It is a different bet on timing and cost. Right now, commercial pressure keeps Coding as the price of admission — no Coding breakthrough, no next round. But in the gaps between Coding battles, the real next chip each house is buying is how much model capacity, data, and compute it is willing to set aside for vision, ahead of time.

One thing is certain: through the second half of 2026, the debate over "native multimodality" is going to keep heating up. Puter's 5 caught visual discrepancies and V4-Flash reaching 1577 on WebDev with just 284B parameters are concrete footnotes in that ongoing argument.

References:

  • 36Kr, "Kimi K3 vs DeepSeek V4: the native-multimodal time gap" (Li Zhaofeng, 2026-08-04): https://36kr.com/p/3924826666301831
  • Kimi K3 release & Arena Frontend Code performance (same source)
  • DeepSeek V4-Flash GA parameters and Arena WebDev score (same source)